The Red Hat AI Inference Server Model Optimization Tools 3.3.5 (CUDA) includes a vulnerability identified as CVE-2025-14926 in the Hugging Face Transformers library. The flaw is due to the convert_config function executing user-supplied strings as Python code without proper validation, leading to potential arbitrary code execution. An attacker can exploit this by supplying a malicious SEW model checkpoint that, when processed by a user, triggers code execution in the user's context. Exploitation requires user interaction and processing of a crafted model file. The vulnerability is rated as important (high severity) by Red Hat. There is no patch or official fix currently available according to the Red Hat advisory RHSA-2026:30078. Mitigation involves avoiding untrusted model checkpoints or using sandbox environments to isolate processing.

Successful exploitation allows arbitrary code execution with the privileges of the user processing the SEW model checkpoint. This can lead to compromise of data integrity, confidentiality, and availability within the scope of the user's permissions. Since administrative privileges are typically not involved, full system compromise is less likely. However, the impact remains high due to the ability to execute unauthorized code.

No official fix or patch is currently available from Red Hat for this vulnerability. Users should avoid converting SEW model checkpoints from untrusted or unverified sources. If processing untrusted models is necessary, it should be done within isolated sandbox environments to limit potential damage. Monitoring and verifying the integrity of model checkpoints before processing is recommended.